30 septembre 2026 à 13h30
Abstract:
An increasing part of research relies on data re-use, meaning it analyzes data that was not primarily collected for research purposes. In this context, an important aspect is that of privacy, as that data may reveal sensitive information that its subjects have not consented to make available.
Over the past five to ten years, synthetic data generation (SDG) has been increasingly discussed, studied and deployed as a means to enable indirect and privacy-preserving data re-use. It consists in learning a generator that approximates the underlying distribution of an observed dataset, then sampling synthetic records from it to be shared for a variety of purposes, including training machine learning models. In order to provide formal privacy guarantees, methods usually rely on the mathematical framework of differential privacy (DP), which is not exclusive to SDG.
In this talk, we provide an overview of the challenges of SDG and main families of existing methods. We also introduce the DP framework, the key research questions associated with it, and exemplify how they apply in the context of SDG.
Paul Andrey (Magnet Team)
ESPRIT building, room ATRIUM
français
English